Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
MP All Models The end of Mac Pro
- Thread starter Mac3Duser
- Start date
- Sort by reaction score
In traditional workstations, memory bandwidth and integrity (out-of-band ECC) are characteristic features.
I would argue with Apple's ARM platforms, Apple blurs the lines. Memory bandwidth is already superb, more than most users (at different tiers) would need. Integrity is helped by on-die ECC in DDR5 and beyond. Since DRAM chips are very closely bonded to the SOC, transfer errors are likely non-existent.
I think the natural step for Apple to implement ECC will be in-band ECC, using part of the capacity (LPDDR being cheap) and bandwidth (in abundance so far). When there is such need, Apple could add it in their future SoC's memory controller.
Worth mentioning Apple just enabled RDMA over Thunderbolt in macOS 26.2.
This enables a programming paradigm for parallel workloads offloading to a cluster of Mac's. Such workloads could be AI inference, 3D rendering & etc.
Now comes an interesting question. When the RDMA paradigm is mature in its ecosystem and if there are enough users willing to pay, a future Mac Pro could be a box of housing multiple boards with each being an separate Mac. All interconnected through RDMA over PCIe or some sort of other physical connectivity appropriate at the time.
Mac Pro in the traditional sense is dead. But don't rule out a new-era Mac Pro.
I would argue with Apple's ARM platforms, Apple blurs the lines. Memory bandwidth is already superb, more than most users (at different tiers) would need. Integrity is helped by on-die ECC in DDR5 and beyond. Since DRAM chips are very closely bonded to the SOC, transfer errors are likely non-existent.
I think the natural step for Apple to implement ECC will be in-band ECC, using part of the capacity (LPDDR being cheap) and bandwidth (in abundance so far). When there is such need, Apple could add it in their future SoC's memory controller.
Worth mentioning Apple just enabled RDMA over Thunderbolt in macOS 26.2.
This enables a programming paradigm for parallel workloads offloading to a cluster of Mac's. Such workloads could be AI inference, 3D rendering & etc.
Now comes an interesting question. When the RDMA paradigm is mature in its ecosystem and if there are enough users willing to pay, a future Mac Pro could be a box of housing multiple boards with each being an separate Mac. All interconnected through RDMA over PCIe or some sort of other physical connectivity appropriate at the time.
Mac Pro in the traditional sense is dead. But don't rule out a new-era Mac Pro.
An M4 Studio Ultra would have been a real game-changer I feel.
Obviously Apple have huge stocks of M3 Ultras, and committed dealers are still shifting these at premium prices. 3.2Ghz is now a difficult sell.
After the huge impact of the M4 4Ghz single-core Minis, an M5 Ultra Studio should be an utter monster - with M6 not long after.
But what do you do as a company when faced with the prospect of your latest tech product making everything in your full warehouse look dated?
I presume the markup and profit is sufficient for them to sit on stuff, sell latest stuff to whoever will pay the premium at the time of their choosing - and eventually shift lesser stuff to locations that are behind the curve? However much posturing there is of ‘smash, bang, wallop’.
Not a criticism, just the law of the jungle.
Obviously Apple have huge stocks of M3 Ultras, and committed dealers are still shifting these at premium prices. 3.2Ghz is now a difficult sell.
After the huge impact of the M4 4Ghz single-core Minis, an M5 Ultra Studio should be an utter monster - with M6 not long after.
But what do you do as a company when faced with the prospect of your latest tech product making everything in your full warehouse look dated?
I presume the markup and profit is sufficient for them to sit on stuff, sell latest stuff to whoever will pay the premium at the time of their choosing - and eventually shift lesser stuff to locations that are behind the curve? However much posturing there is of ‘smash, bang, wallop’.
Not a criticism, just the law of the jungle.
Last edited:
I don't really like the Mac Studio. It's mainly because it's not easy to open and it doesn't have multiple NVMe ports that can be swapped out. For me, it's not a workstation.
Unified memory, okay; AI on a large amount of RAM, okay; I can see people saying that a Mac Studio can compete with the Nvidia DGX or an RTX 6000 Blackwell 96GB, but for me, that's not how it works.
A workstation is a motherboard, a chipset, and a processor with a lot of PCIe lanes. I'm repeating myself a bit, but that's what it's about. PCIe lanes and PCIe slots. Trying to route everything through Thunderbolt and external enclosures isn't a good solution for me. I prefer a solution is a case with a large power supply that powers everything inside, and cards that can be swapped out depending on the task. So, PCIe cards with NVMe drives, Ethernet cards, graphics cards, cards for musicians, and so on...
Indeed, if we can have a case with two M5 Ultra motherboards to create an internal cluster powered by a single power supply and driving multiple NVMe slots, we can imagine a new type of machine. That's becoming interesting, but I'd be more inclined to want a macOS that remains open to Intel (and AMD) and accepts AMD and Nvidia drivers.
Unified memory, okay; AI on a large amount of RAM, okay; I can see people saying that a Mac Studio can compete with the Nvidia DGX or an RTX 6000 Blackwell 96GB, but for me, that's not how it works.
A workstation is a motherboard, a chipset, and a processor with a lot of PCIe lanes. I'm repeating myself a bit, but that's what it's about. PCIe lanes and PCIe slots. Trying to route everything through Thunderbolt and external enclosures isn't a good solution for me. I prefer a solution is a case with a large power supply that powers everything inside, and cards that can be swapped out depending on the task. So, PCIe cards with NVMe drives, Ethernet cards, graphics cards, cards for musicians, and so on...
Indeed, if we can have a case with two M5 Ultra motherboards to create an internal cluster powered by a single power supply and driving multiple NVMe slots, we can imagine a new type of machine. That's becoming interesting, but I'd be more inclined to want a macOS that remains open to Intel (and AMD) and accepts AMD and Nvidia drivers.
Good points there.
Except that this new desktop beast would surely have silent liquid-cooling, and a built-in UPS, with an external battery?
Old stereotypes die hard. Especially true for folks like tinkering. I think such folks will be better served by x86_64 workstations. That's for task/project specific workload. And get a Mac mini as sidecar for admin/email/web/paperwork.
MacOS is sinking in quality and usability and loses focus from Apple. While it's sinking, perhaps so far still better than alternatives in terms of UI and cleanness.
If Apple can figure out the 4x combined chip ‘Extreme’ that was planned, during one of these next M generations, i bet they will make it Mac Pro only
I don't know. I haven't installed Tahoe on my Macs; I don't like Liquid Glass at all. The ergonomics are heading in a direction I don't like. I'll probably buy a MacBook Pro M6 next year, though. But no desktop Macs.
And I'll minimize my use of Liquid Glass as much as possible.
For my workstation, as I said, I'll definitely get something from another brand, where I can install Linux (and Windows if needed for a few apps I require).
I'm rather opposed to cloud solutions, and I think there will always be solutions for local storage and local computing, and I'll always be a customer of local computing.
Tahoe is actually sluggish even on my super “exotic” M2 MBA, so it’s not just lagging on Intel powered machines. I suppose I need to upgrade to M4 Max MBP to get acceptable speed with all those glass effects.
Every time I click the Applications fly out it’s like a domino effect the app icons appearing.
Away from that, my Intel Mac Pros are still quick however.
Tahoe is actually sluggish even on my super “exotic” M2 MBA, so it’s not just lagging on Intel powered machines. I suppose I need to upgrade to M4 Max MBP to get acceptable speed with all those glass effects.
Every time I click the Applications fly out it’s like a domino effect the app icons appearing.
Which is a bit sad considering that's instant (at most 2 frames) on Mac Mini 2018 i3 w/8GB of RAM running Mojave.
There's relatively few workstation-class things those machines can't do when properly programmed. Unfortunately, Apple has somehow managed to add lag to things that were fast enough in the G4 days (like opening Applications or navigating System Settings).
The Mac Pro is now an odd beast, I really struggle to see its purpose except in maybe some highly specialised roles that are few and far between. Apple Silicon is just not designed with user expansion in mind, the GPU and memory are all in the chip!
Maybe if you could get additional PCI compute modules and turn the Mac Pro into a self contained AI cluster that might be enticing to a few people, but even then its a hard sell against many other PC based AI machines that will be much cheaper.
But I truly feel the OP's pain. For me, number one reason why I would go Mac Pro is because the only alternative (Mac Studio) is possibly Apple's most unoriginal and boring design ever. Just a square grey box with some ugly ports on it. There was literally no imagination that went into the exterior design of that thing. It was like someone challenged the design team to put people to sleep as fast as possible. This is the real issue for me, Apple Silicon is genuinely incredible and deserves to consign the Mac Pro to the history books, but it should also go hand in hand with a decent enclosure.
Bring back the trash can design, those Mac Pro's looked awesome and its hard to imagine that the thermal limitation of the case will be much of a problem for AS!
Maybe if you could get additional PCI compute modules and turn the Mac Pro into a self contained AI cluster that might be enticing to a few people, but even then its a hard sell against many other PC based AI machines that will be much cheaper.
But I truly feel the OP's pain. For me, number one reason why I would go Mac Pro is because the only alternative (Mac Studio) is possibly Apple's most unoriginal and boring design ever. Just a square grey box with some ugly ports on it. There was literally no imagination that went into the exterior design of that thing. It was like someone challenged the design team to put people to sleep as fast as possible. This is the real issue for me, Apple Silicon is genuinely incredible and deserves to consign the Mac Pro to the history books, but it should also go hand in hand with a decent enclosure.
Bring back the trash can design, those Mac Pro's looked awesome and its hard to imagine that the thermal limitation of the case will be much of a problem for AS!
Last edited:
Y'all remember the NeXT cube? Backplane with 4 slots - 1 occupied by what we would call the motherboard. You could populate the other 3 slots with massively powerful (for the time) specialized co-processors. NeXT Dimension IIRC was the video accelerator, there was a DSP board, and maybe something else - y'all will correct me 🤣.
To me, the king of workstations that would actually warrant a $6k starting price would look like this in MP form factor:
- dual socket intel
- massive RAM ceiling
- Apple Silicon PCIe co-processor cards
Why would you buy anything else? Pipe dream, I know.
Beautiful design. The same for a new Mac Pro would be wonderful.
Last edited:
It's really not hard: it's mainly for people who need specialised I/O cards, particularly for audio/video work, for which internal PCIe4 offers more bandwidth-per-card than Thunderbolt. Things like ProTools audio cards. In the past that has been a significant part of the "Pro" Mac market. It's probably a shrinking pool as Thunderbolt bandwidth increases and new Thunderbolt-based hardware products appear.
That might have been the other way to take the Mac Pro - using something like MPX cards as compute modules - but the host would either need to stick with x86 (not happening) or Apple would have to make a new Apple Silicon die just to run all of those PCIe lanes. It would also be very, very dependent on software being optimised for that kind of set-up. The concept could be proven with Thunderbolt-linked Mac Studios.
The Mac Studio isn't perfect but it's doing its job - powering the system-on-a-chip, keeping it cool, and providing sockets for the interfaces and otherwise keeping out of the way. It's neat, quiet, can be self-contained (most SFF PCs have a huge power brick and noisy fans) and blissfully free of rainbow "unicorn vomit" LEDs. Feel free to add your own go-faster stripes.
Now, if we're talking subjective aesthetics I'd say the "classic" Mac Pro tower was a nice design - but the 2019 Mac Pro is, IMHO, a grotesque bit of over-engineering and tacky "steampunk" design that feels like a caricature of the original cheesegrater - and the 2023 version removes any vestage of justification for such a design.
My two concerns with the Mac Studio are (a) dust accumulation and (b) the use of proprietary, non (officially) upgradeable SSD modules rather than M.2 NVMe sticks. Unified, on-package RAM may be more efficient, but I need convincing that the advantage of on-die NVME controller + proprietary Flash-only modules outweighs the advantage of using standard M.2 NVMes.
Would you like a hockey-puck mouse and a butterfly keyboard to go with that?
The whole "trashcan" design was built around that three-way cooling system designed for a specific CPU + 2xdGPU chipset. Maybe you'd get away with it with Apple Silicon, but the Mx Ultra can still whack out heat and relies on a big chunk of purpose-designed copper in the Mac Studio so I wouldn't count on it. In any case, an Apple Silicon mainboard would only use 1/3 of the case... and it's not as if the Trashcan offered any real upgradeability beyond the RAM.
Noooo....! that can't be true, Steve Jobs hated upgradeability.
Problem is, everything that came on those expansion cards is now built into the Apple Silicon package, you don't need a MO drive the size of two housebricks,, or hunks of spinning rust. The power supply (between more efficient chips and GaN PSU tech) is a fraction of the size so, I dunno, maybe you could fit in a dubstep-grade subwoofer, or a beer cooler or something to use up all of that empty space?
Also, the fastest external interface available at the time was probably parallel SCSI at, what, a few hundred kB per second? One of the reasons the internal slots were a must back then.
(Intel meaning x86) Not happening. Apple have bet the farm on Apple Silicon which works great in the laptops, mobiles and SFF devices that make Apple money. Intel Mac is dead.
Is that RAM ceiling on the host processor(s) or the co-processor cards? The RAM ceiling comes from using on-package Unified, LPDDR RAM with high bandwidth - which is also one of Apple Silicon's efficiency advantages. Or, you could get a bunch of Mac Studios (which will always be much cheaper than small-volume PCIe compute modules) and Thunderbolt them together.
It was a beautiful design for an early-90s 68k box that needed humungous expansion cards (with a 32-bit parallel bus) for stuff that an iPad can do on a single chip today.
Maybe you don't understand my words. Just as the design of the trashcan could be adapted to a Mac Pro AS (more storages in the place of gpus), the design of the Cube Next could also be adapted: we want to be able to add cards, you know, SSDs, all sorts of things. And if this black cube has to be smaller than the Next, no problem. We want to be able to open it. To clean the dust and to add the NVMe storages too. And coprocessor cards yes maybe. Why not ?
(Sharp X68000 design was beautiful too)
Do you realise only 3 (THREE) years unofficial support left from Apple at most ?
No more MacOS x86_64 binaries will be released by Apple after support of MacOS 26 ends in three years.
Internally though I'm sure Apple will continue to compile MacOS for x86_64 in the foreseeable future.
LLMs require GPUs access to large amount of DRAMs. Apple's ARM platforms are cheaper alternatives atm when compared to x86_64 platforms built with professional GPUs from Nvidia or AMD.
While I believe AI on the cloud will remain the de facto choice for many people, local AI i.e. running LLMs inference on hardware hosted at home or in office will be significantly large, simply out of the demand of privacy and trade secrets.
So Apple may release a new-era of Mac Pro which is a 'cluster of Macs' inside a box with very efficient interconnect and cheap SoCs (with defects that are destined for landfill otherwise). E.g. Other than the 'main Mac' inside the box, other 'boards of Mac' could re-use SoCs with unacceptable defects in CPUs but with good to perfect GPUs.
Airgapped PC for work and Apple for all the rest is how I've always handled it. Mac OS is pleasant to use and not a focus of malware so using it on the web seems ideal, really. Unlike Winblows.
One thing that I had happen several times during work is the power supply going out after a few years of regular heavy loads. Also one borked mainboard once. With a Mac Pro I'd be hosed, no - everything proprietary and the 'geniuses' at the help desk not exactly able to help right away and needing to escalate to the guys who know more than how to factory-reset an iPhone?
With a PC all it took was a quick trip to the nearest computer store and I was up and running again in a matter of hours - would have been quicker if it didn't require time to diagnose the problem first. Same for storage, memory, GPU I reckon.
One thing that I had happen several times during work is the power supply going out after a few years of regular heavy loads. Also one borked mainboard once. With a Mac Pro I'd be hosed, no - everything proprietary and the 'geniuses' at the help desk not exactly able to help right away and needing to escalate to the guys who know more than how to factory-reset an iPhone?
With a PC all it took was a quick trip to the nearest computer store and I was up and running again in a matter of hours - would have been quicker if it didn't require time to diagnose the problem first. Same for storage, memory, GPU I reckon.
Last edited:
Interesting discussion.
For me, being purely a music-maker, with a single 1440p 21/9 monitor, and a couple of M2-drives on a powered hub - even the base Mac Mini M4 is overkill when it comes to GPU. I like it’s size and quiet operation though, and the screen image is nice. Very stable.
I’d certainly be happy with being able to add just one AES/ABU audio output to my Mini - without having to spend $2000 on a converter box.
Something a little more modular would be refreshing perhaps, but I really think there could/should be an M5 Ultra Music Studio model, that has only 12 GPU cores, with S/PDIF or AES/ABU outputs for external DACS.
For me, being purely a music-maker, with a single 1440p 21/9 monitor, and a couple of M2-drives on a powered hub - even the base Mac Mini M4 is overkill when it comes to GPU. I like it’s size and quiet operation though, and the screen image is nice. Very stable.
I’d certainly be happy with being able to add just one AES/ABU audio output to my Mini - without having to spend $2000 on a converter box.
Something a little more modular would be refreshing perhaps, but I really think there could/should be an M5 Ultra Music Studio model, that has only 12 GPU cores, with S/PDIF or AES/ABU outputs for external DACS.
In four years, the new MacPro will be worth... 20% of what you paid for it. Hackintosh will also lose value, but you will pay four times less for it. You have to be able to count. A computer is not an investment, but a loss spread over years. But you can drastically reduce costs or pay only for the brand.
I’m not sure that’s entirely accurate. At the consumer level, where you might want to run office applications but also run inference on a local model then the unified memory architecture that Apple offers does give an edge at a competitive price point (vs a top end GPU bundled with all the other bits and pieces you need to get a PC to work). This is not exclusive to Apple though and nothing to do with ARM64, AMD has the Ryzen 395 that also uses UMA and is X86 64 bit.
You can also get some reasonable performance when clustering the M4 Mac Mini, but as soon as you start clustering with Apple the cost v performance quickly falls away as the network is a serious bottleneck, hence my thoughts on using the PCI bus which would give a much better throughput and lower latency. But you will still be limited by the architecture, specially the M2 Ultra is not designed to support external GPU, even just for compute. This is maybe something they could address with a processor upgrade, but like I said before, it’s going to be a very expensive system, first the Mac Pro is not cheap and presumably bespoke PCI compute modules designed for it would also have the Apple tax applied to them.
So yes, for local inference and maybe some limited training a Mac Studio or even a MacBook Pro with a max processor will beat a similarly priced PC/GPU in terms of raw performance (but I still think NVIDIA are going to win on the software stack!!), but for anything more where you start clustering machines you are almost certainly going to get more bang for your buck outside the Apple ecosystem.
Hate to break it to you but the Next Cube was neither the first or last computer to let you add cards or storage, and the cards it used were a completely different form factor to the PCIe cards that people want today and ISTR that the storage was a brick-sized Magneto-Optical drive which was about as unlike a SSD as you could get. As for the Trashcan: the only upgradeable thing about that was the RAM - it had a single, proprietary SSD card and - just like the Studio - you were expected to use external Thunderbolt devices for all expansion.
If by "adapt" you mean make something a different size with different cooling requirements and taking radically different-sized components and then wonder why it's now a totally different shape - have fun.
Apple is never going to make another device that can have another company's hardware physically installed inside it. The reason given will be "security".